THE QUARTERLY JOURNAL OF EXPERIMENTAL PSYCHOLOGY, rWL44A QrlyS-210
A Reinterpretationof SomeEarHerEvidencefor
of Implicitly AcquiredKnowledge
Abstractiveness
Pierre Perruchet
Paris, France
UniversitéRené-Descartes,
JorgeGallego
Université Pierre et Marie Curie, Paris, France
Chantal Pacteau
Université René-Descartes, Paris, France
Reberand Lewis (1977)exposedsubjectsto a subsetof letter stringsgenerated
from a synthetic grammar, then asked them to reorder scrambledletters to
generatenew grammatical strings. The distribution of the frequencyof the
bigramscomposingtheir solutionscorrelatedbetter with the frequencyof the
bigramscomposingthe full set of stringsgeneratedby the grammar than with
the frequencyof the bigramscomposingthe subsetof strings displayedin the
study phase.In his recentoverviewon implicit learning,Reber(1989)develops
this eiperimental result into one of the main supportsfor his contention that
studying grammaticalletter stringsgivesaccessto the abstractstructureof the
grammar.
However,this result can be accountedfor by a set of biasesinherent to the
Reber and Lewis procedure.In the presentexperiment,a group of subjects
learned from a list of the bigrams making up the study strings, a condition
which precludes the abstraction of any high-level rules. The pattern of
correlations outlined above also emerged in this condition, thus lending
support to our re-interpretation.
The notion of implicit learning originates from the field of artificial grammar
learning, as explored over the last twenty years by Reber and his associates
(e.g. Reber, 1967; Reber, Allen, & Regan, 1985; see Reber, 1989, for a
review). In a typical experiment, subjects are first exposed to a set of letter
strings generated from a synthetic grammar that defines authorized letters
and the permissible transitions between them. Subjects are given no informa-
Requestsfor reprints should be sent to P. Pemrchet, Laboratoire de PsychologieDifférentielle, 28 rue Serpente,75006PARIS, France.
O 1992The Experimental Psychology Society
PACTEAU
GALLEGO,
194 PERRUCHET,
tion about the rule-governednature of the stimuli and are askedto perform a
task that diverts them from the searchfor regularitiessuch as rote learning
(e.g.Reber,Kassin,Lewis,& Cantor, 1980),short-termmatching(Mathews
èt àt., 1989),or liking judgement(McAndrews& Moscovitch,1985)of letter
strings. Subjectsare subsequentlyasked to categorizenew letter strings as
grammatical or non-grammatical;non-grammaticalitems are formed from
ih. ru-e subsetof letters, but violate transition rules. Resultsconsistently
show that subjectsperform this categorizationtask better than chance.As
they are apparentlyunableto verbalizethe rulesunderlying their decisionof
well-formedness,their performanceis thought to be the end-product of an
unconsciousabstractionprocess.
In recentyears,thesefindings have receivedfurther support from studies
basedon other paradigms.For instance,Lewicki, Hill, and Bizot (1988)
studied the ability of subjects to improve their performance when the
location of a responsesignalwasdeterminedby the pattern of its location on
specificearlier trials. After extendedpractice, subjectsexhibit a substantial
d.rrr"r. in reaction time to target signals with predictable locations,
although they are unableto articulate any of the complexrules that regulate
the sequenceof trials (cf. also Lewicki, Czyzewsk&,& Hoffman, 1987;
& Bullemer,1987).Implicit
McKetvie,1987;Millward & Reber,1972;Nissen
process
control tasks,in which
learning has beenalso shown for interactive
he
or she receives.For
the learner influencesthe nature of the stimulation
subjectswere
(1983)
experiments,
instance,in one of Berry and Broadbent's
transportation
city
of
a
required to imagine that they were in charge
department and \ryereinstructed to manoeuvre the time interval between
tWith
buies to reach and maintain a specifiedload of passengersper bus.
training in a computer-simulatedinteractive situation, subjectswere able to
make appropriate adjustments,although they were unable to verbalizethe
actual function relating the two variables(cf. also Broadbent,Fitzgerald, &
Broadbent, 1986;Hayes & Broadbent, 1988;Stanley,Mathews, Buss, &
Kotler-Cope,1989).On the whole,thesestudiesapparentlyprovideconsistent evidencethat peoplecan unconsciouslyabstract environmentalregularities and use this tacit knowledgeto improve performance.This ability is
consideredto be highly relevant to accountsof how humanscope with the
complex environmentof everydaylife and is thought to encompassperceptual and socialbehaviour,language,and so on.
Another set of studies suggests,however, that the knowledge base
underlying improvementin performancein implicit learning situationsmay
As noted by Broadbent (in press),someof
be availableto consciousness.
betweenstudiescouldbe dueto the methodof operationalthe discrepancies
ization of consciousness.Typically, authors positing that some knowledge
always remains inaccessibleto conscious reflection, such as Reber and
IMPLICITLEARNINGAND ABSTRACTIVENESS195
Lewicki, rely on recall-likeverbal protocols, whereasauthors who challenge
this claim use recognition-liketests(e.g.Dulany, carlson, & Dewey, lgga;
Perruchet & Pacteau, 1990).Another issue ffiây, however, be of greater
importance. Authors stressingthe availability to consciousnessof the
knowledgeunderlying performancealso claim that this knowledgeis far less
complexand abstractthan is commonlythought.
In artificial grammar learning, severalstudiesindicate that specificand
fragmentary knowledgeof study items components,in particular first and
last letters and permissiblebigrams (i.e. pairs of letters) or trigrams, is
sufficient to account for grammaticarityjudgements (Dulany et à1., l9g4;
Mathews et al., 1989;Perruchet& Pacteau,1990).Perruchetand pacteau
(1990)obtainedexplicit knowledgeon valid bigramsthrough a recognition
testand then performedquantitativesimulationsof grammaiicaljudgéments
on the basisof thesedata. The resultingperformancepattern matcheànicely
with observedresults when simulatedjudgementsôf non-grammaticality
weregeneratedfor all teststringscontainingat leastoneunrecognized
pair of
letters.In a more sophisticated
way, Druhan and Mathews(1989)devéloped
a computationalmodel basedon Holland et al.'s (1986)induction theôry.
They usedverbal instructionsgeneratedby subjectsstudying letter stringsin
standardconditionsas input. When the rules derivedfroÀ verbal instructions wereautomaticallytunedby an optimizationalgorithmusingfeedback
on correctnessof responses,simulatedperfonnanceapproachedthat of the
original subjects.
Similar conclusionsmay be drawn from studiesinvolving other implicit
learningparadigms.Perruchet,Gallego,and savy (1990)observedthat the
rulesconstitutingthe seriesof trials in the Lewicki et al. (1988)paradigmalso
alteredthe relativefrequencyof particulartransitionson the whole sequence
of trials. They provide evidence,through replicationand re-analysisbf the
Lewicki et al. data, that subjects'perfonnancemay simply reflect this far
more elementaryinformationalcontent.Likewise,Sanderion(1989)showed
that perfonnancein interactivecontrol taskscould be due to fàmifiarization
with peripheralaspectsof the task rather than to the internalizationof the
underlyingstructureof the situation.
In fact, theserecent studiesdo not demonstratethat people really use
-ny
consciousknowledgeof surfacefeaturesof the situation.
showingthat
this knowledgecould be sufficientto account for perfor-unôr, they merely
providean alternativeinterpretationof the traditionalabstractionistpoint
of
view. Comparingthe explanatorypower of thesetwo interpretationsclearly
requires other experimentalevidence.In his recent overview on implicit
learning,Reber(1989,pp.22sff) puts forward two main argumentsto
show
that implicit learninggivesaccessto deep,abstractstructureof the stimuli,
in
addition to fragmentarypiecesof knowledgeon their surfacefeatures.The
pAcrEAU
196 pERRUcHET,
GALLEGo,
presentpaper is aimed at examiningthe empirical and theoreticalvalidity of
thesetwo arguments.Becauseboth refer to the field of grammarlearning,the
following discussionwill be limited to this area of research.
Reber's First Argument:
The Transfer to Superficially Different,
but StructurallySimilar Stimuli
The first argumentdraws on resultsinitiatly collectedby Reber(1969),and
morerecentlyreplicatedand extendedby Mathewset al. (1939).Briefly,these
studies demonstrate that learning from letter strings generated by one
grammar transfersto letter stringsgeneratedby a grammar sharingthe same
formal rules, but using another set of letters.Theseresultsare interpretedas
showingthat subjectsare able to abstractautomaticallythe "syntax" of the
displayedmaterial, and they can usethis structural knowledgeindependently
of the "vocabulary".
Although Mathewset al. (1989)adhereto Reber'sabstractionistposition,
certain featuresof their resultsrun counter to a strict versionof this point of
view. For instance,transferto a different letter setis better under intentional
than under implicit instructions,whereasthe theory posits unconsciousness
of abstraction.Furthermore,transferto a different letter setoperatesequally
well for yoked subjectswho learn from verbal instructions given by experimental subjectsas for the experimentalsubjectsthemselves.Most of these
verbal instructions are to select or avoid specific bigrams or trigrams in
specificlocationsin the strings.Hence,transfercan hardly be put forward as
an argument for an abstraction processgenerating high-level rules from
entire letter strings.
The results of Mathews et al. leave open one possibility that people
abstract"local syntaxes".However,eventhis weak versionof the abstractionist viewpoint may be challenged. On formal grounds, the artificial
grammar setting is a categorizationtask. Researchconductedover the last
decadeor so in the field of categorizationhasconvincinglydemonstratedthat
the capacity to transfer to new test items does not prove that people have
elaboratedabstractrepresentations
from the studyexamplars.This hingeson
two core issues. The first pertains to the moment of processing and
discriminatesbetweenwhat Estes(1986)has termed "late" versus"early
computation" models.In early computationmodels,processesunderlying
categorization are accomplishedduring the study phase, and their end
products are merely examinedat the test time. In late computation models,
studied examplarsare thought to be stored in memory without transformation, and all the computationsneededfor making a decisionare performed
when the subjects encounter new items. The postulate of Reber and
Matthewset al., that learningis implicit, is obviouslyat variancewith late
computation models,which have receivedincreasingsupport in the current
IMPLICIT
197
LEARNING
ANDABSTRACTIVENESS
literature. The second issue concerns the mode of processing. In the
traditional view, people generateabstract representations,such as schemas
or prototypes. More recent interpretations, inspired by seminal papers by
Medin and Schaffer (1978) and Brooks (1978), focus on non-analytic,
analogicalprocessing(for a wider-rangingoverviewoseeJacoby & Brooks,
1984;Medin & Ross,1990).
To illustrate how theserecentdevelopmentsapply to the field of grammar
learning, assumethat subjectshave been asked to learn trigram CVC, and
then exhibit positive transfer with trigram DJD. In the traditional view,
transfer is taken as evidencethat subjectsautomatically use the repetitive
presentationof CVC to abstract somethinglike "one letter surroundedby
two identical letters", then store this formal rule in memory and apply it
when coping with a new letter string. An alternative explanation is that
subjectsstore the specifictrigram CVC, and deal subsequentlywith DJD in
the sameway as they dealt with CVC on the basis of the global similarity
betweenboth items.The key point hereis that this alternativeinterpretation
accounts for transfer without assuming that people form an abstract
representationwhile studying letter strings (for further discussionon these
points, seeMathews,in press;Perruchet& Pacteau,l99l).
Reber's Second Argument:
The Performance Pattern Reflects the Underlying Grammar,
Not the studiedstimuli
Reber's(1989)secondargumentis basedon an intriguing experimentalresult
initially reported by Reber and Lewis (1977).In this study, the standard
categorizationtest was replacedby a task in which subjectshad to reorder
scrambled letters to generategrammatical strings. Nature of knowledge
resulting from exposureto grammaticalletter strings is assessed
by detailed
analysis of performance on this task. The bigrams occurring in subjects'
solutions to the anagrams were tabulated, and their frequencieswere
comparedagainst(a) the frequencyof occurrenceof the bigramscomposing
the strings actually displayedin the study phase,and (b) the frequencyof
occurrenceof the bigramscomposingthe full set of stringsgeneratedby the
formal grammar (hereafter:the virtual strings).The correlation coefficients
were0.04and 0.72respectively.
Reber(1989)concludesthat the failure of the
correlation involving the study phase bigrams "to be different from zero
suggeststhat subjects \ryerenot solving the anagrams on the basis of
superficialknowledgeof frequencyof bigrams". Rather, the high correlation
with the bigramsmaking up the virtual stringsindicatesthat subjects"clearly
acquiredknowledgethat can be characterizedas deep,abstract,and representativeof the structureinherentin the underlying invariancepatternsof the
stimulusenvironment"(Reber,1989,p.226).
198 PERRUcHET,
GALLEGo,
PAcTEAU
Although the rationale for this argument is sound, a number
of biases
potentially flaw the empiricalfindingson which it is grounded.
Let us start by
considering the selection of the study strings fùm the virtual
strings.
Selectionhas two direct consequences.
The firit is to reducedrastically the
variability of bigram frequency. The variance of the distribution
of the
bigram frequencyof selecteditems is about one ninth of the corresponding
valuecalculatedfrom the whole set(3.85vs. 35.89).Low variance
oùviously
makesit difficult to obtain high correlations.The secondeffectof
selectionis
that the distribution of frequenciesof the bigramsmaking up the
study and
the virtual itemsdiffers.This constitutesu n...rr"ry condition for
obtaining
different correlations with both sets of data, as Reber and Lewis
did.
However,closescrutiny of the material showsthat a bias may
have been
introduced.Considerthe under-represented
bigrams-that is, thosebigrams
havinga lower proportion in the study itemstÉan in the whole
sample.The
observedpattern of correlationsshowi that thesebigramsare especiâily
easy
to learn; indeed, the higher correlation of subject generatedbigrams
with
virtual than with studyitemsimpliesthat the rrndrr-rrpresented
bigru-, up
generatedin anagramsmore often than could havebeenpredicted
iro- their
effective occurrencesin studied items (the reverse is tnre for
the overrepresented
bigrams,which are not dealtwith herefor simplicity'ssake).The
four under-representedbigrams exhibiting the largest differences
between
studyand virtual frequencies
are TV, TT, Vv, and xx. TV is an extremely
common abbreviation,and other bigramsare doublets.These
bigramsmay
have been especiallyeasy to learn becauseof their relative salience
rather
than becauseof their frequencyin the virtual set of grammatical
strings.
Another more damaging bias than those deriving from
study item
selectionarisesfrom choiceof the testmaterial.Subjectsitartfrom
stringsof
lettersto solveanagrams,which placessevereconstraintson the
natureof the
gairs that may be produced.Supposesubjectslearn in the study phasethat
vv is a valid bigram. They have littie means of demonrtàting
this
knowledgeif thereare few Vs in the testmaterial,and overall,
the frequency
of VV in their responsewill be partly dependenton the number
of available
Vs' In the Reberand Lewisstudy,the letier stringsusedin the
anagramtask
Y.eteobtainedby scramblingthe grammaticalitems that were nJt initi"Uy
displayed.As a consequence,
su.bjectsare prompted to fit their bigram
productionwith the bigramsmaking up thesegrammatical
items.Indeed,if
-or,
theseitems containeda large numbir
ruy,-vv bigrams,subjectshad a
large number of vs at their disposal in tne ,rrurnbl.d
strings, which
facilitatedVV production.The crucialpoint is that the frequenry
oîbigrams
composingthe grammatical test strings before scramblini ctosety
matched
the bigram frequencyin the virtual strings (the two distributions
correlated
to 0'949) and were only moderatelyrelaied to the bigram frequency
in the
study strings(r=0.234). This correlationalpattern is iot simply
an unfortu-
199
LEARNING
ANDABSTRACTIVENESS
IMPLICIT
nate chanceeffecqbut, rather, reflectsa logical constraint. The fact that the
virtual stringswereexhaustivelypartitioned into study and test stringsmakes
the frequencyof bigramscomposingthe virtual stringsthe (weighted)mean
of bigramscomposingstudy and test strings.It can be demonstratedthat the
correlation betweentwo given seriesof values(at leastwhen this correlation
is positive) is lower than the correlation betweenone of the seriesand the
mean of the two series.Thus, to sum up, the constraintscreatedby the test
material make it impossiblefor subjectsto produce bigrams mimicking the
distribution of bigrams in the study strings, whereas they can almost
perfectly parallel the distribution of bigrams in the virtual strings.
Overview of the Present Study
The experiment describedin this article was designedto strengthen our
contention that the Reberand Lewis (1977)correlationalresult cannot serve
as evidencefor Reber's(1989)abstractionistinterpretation,by showingthat
an identical pattern of resultscan be obtained when subjectsare not given
information neededfor abstractinghigh-levelrules.
The procedurewasthe following: The first group of subjects(string group)
studiedgrammaticalstringsof lettersin the study phaseand werethen given
an anagramtask in the test phase.The material used in both phaseswas
identicalto the one usedin the original Reberand Lewis (1977)study.The
string group was designedto replicate the original correlational results.
However,our proceduredifferedfrom Reberand Lewis in severalways,in
particular in the amount of training given. In the original study, subjects
were exposedto four learning sessions.This has at least two negative
First, the study items are displayed before the anagram
consequences:
Thus, subjectshavea clearidea of the
solvingtrials on eachof the sessions.
task requirementswhen dealingwith the relevantinformation and may have
engagedin explicitmodesof learning,includinghypothesisformulation,rule
testing,mnemonicstrategies,and so on. The secondtouchesmore specifically on the correlational results. As subjects progressivelysolve more
anagramscorrectly, correlations between bigrams frequenciesapproach
fixed values,which dependexclusivelyon experimentalmaterial.If subjects
solve all anagramscorrectly, the correlationsof the frequencyof bigrams
making up the anagram solutions (which are, in this case,the subsetof
grammaticalletter strings not initially displayed)with, for instance,the
bigramscomposingvirtual stringsis 0.949,as statedabove.To correct for
theseproblems,duration of learningwas restrictedto a singlesession(the
power of inferential tests \ryasensuredby collecting data on a far larger
sampleof subjectsthan Reber and Lewis did).
The secondgroup of subjects(bigram group) was givenin the study phase
the individual bigrams making up the strings displayedto the string group.
pAcrEAU
200 pERRUcHET,
GALLEGo,
As in Perruchetand Pateau (1990,expt.l), this was intended to convey
information on valid bigrams and their frequency of occurrence,but to
precludethe abstractionof any high-levelrules. The test phaseconsistedof
the sameanagramtask as for the string-groupsubjects.The mean scoresfor
the bigram group were expectedto be much lower. In the abstractionist
framework, this would occur becausesubjects cannot abstract the deep
structure of the grammar during the study phase.The studiesshowing that
subjectsexposedto letter stringsacquiredspecificknowledgepertaining, for
instance,to the first or last valid letters(e.g.Mathewset al., l9S9) provide
another interpretation. Thus bigram-group subjects,who only observed
letter pairs in study phase,could not acquire this knowledge,which is
obviously useful for solving anagrams.The correlational resultsare of more
fundamentalinterest,in that they discriminateconcurrentinterpretations.In
keeping with Reber's hypothesis,there is no reason to expect that the
frequencyof bigramscomposingsubjects'generatedstringsshould correlate
better with the frequencyof the bigramscomposingvirtual stringsthan with
the frequencyof bigramscomposingstudy strings. However, if the correlation pattern is due to one or a combinationof severalof the biasespointed
out above,strongercorrelationswith bigramscomposingvirtual than study
strings must emergein both the bigram group and the string group.
Method
Subiects. Subjectswere 53 first-year university students majoring in
psychology.Twelve additional subjectsparticipated to the experiment,but
they were subsequentlyeliminatedfrom data analysisbecausethey failed to
follow instructionson the anagramtasks.Thesesubjectsgeneratedat leastl0
anagramswherethe number of constituentlettersdid not match the number
of lettersin the original strings.
subjects were run in 3 groups of 21, 21, and 23. They \ryererandomly
allocated on an alternating basis to experimentalconditions within test
groups. To make this possible,the instructions specificto eachgroup were
provided on separatesheetsof paper.
Materials. The stimuli consistedof letter stringsgeneratedby the Reber
and Lewis (1977)grammar shown in Figure l. The study items for the S
group were the 15 letter strings usedby Reber and Lewis. The strings were
typed on a sheetof paper in a singlecolumn in random order (seeTable l).
The bigram group was shown the 78 bigrams that can be extractedfrom
the 15 letter strings.Theserveretyped in six columns,in random order, as
displayedin Table 2.
The testphaseusedthe 28 remainingletter stringsthat the grammarcould
generate.Theseletter stringswere scrambledand presentedin two columns
in random order (seeTable 3). A dotted line was provided opposite each
letter string for the subject'sresponse.
201
ANDABSTRACTIVENESS
LEARNING
IMPLICIT
FIG. l. Schematicdiagram of the grammarusedin the presentexperiment(taken from Reber
& Lewis, 1977).
TABLE1
List of Letter Strings Displayedin the Study Phaseto the String Group
PTTTVPS
TSSSXS
PVV
PVPXVPS
TSSXXVV
TXXVPXVV
PTVPS
PVPXVV
TSSSSSXS
PTTTVV
TSXXVPS
TXS
TSXS
PVPXTVPS
PTVPXTVV
Procedure
For all subjects,the sessioncomprised a learning phaseand a test phase.
Subjectsfirst receivedtwo booklets in a folder. They rwereorally asked to
removethe first booklet from the folder and to read the instructionsprinted
on the front page.Instructions for the string group (n:26) and the bigram
group (n:27) \ileresimilar, exceptfor variations adaptedto the nature of the
displayedmaterial:
PACTEAU
GALLEGO,
2O2 PERRUCHET,
TABLE2
in the StudyPhaseto the BigramGroup
Listof BigramsDisplayed
xs
xT
sx
vv
vv
vP
VP
SS
XV
TT
xx
Px
TS
PX
PV
TV
SX
SS
TS
PT
VP
SS
ss
Ps
xv
Px
VP
TV
PS
PX
PT
TX
VV
SX
XS
VP
XT
PV
TV
VP
PS
TT
TS
VP
TX
XV
xv
VV
Tv
TT
sx
VP
PT
TV
VP
xx
TT
SS
Pv
XV
vv
XX
VV
VP
rv
SS
PX
TS
PT
XS
SS
xv
Ps
xs
PS
VP
TS
sx
TABLE3
in the TestPhase
LetterstringsDisplayed
Listof Scrambled
XPPVVTV
VPTPST
PVVXVTPT
SSTSXSS
TXPTPVVV
vxxTPssT
TXTVVSSX
PXVTPVPS
TTTXXVSV
TTPTTVPS
SPXTVX
vxvTx
TVSPXTX
TTVXVX
(PTVPXVV)
(PTTVPS)
(PVPXTTVV),
(TSSSSXS)
(PTTVPXVV)"
(TSXXTVPS)
(TSSXXTVV)
(PTVPXVPS)b
(TSXXTTVV)
(PTTTTVPS)
(rxxvPs)
(TXXVV)
(T XX T VP S)
(T Xx T Vv )
xsvvrssx
VT P TV TTT
xvxvTs
STXSS
VTVTTXX
VTPV
VXTTTSXP
VXXSSTPS
XPVTVVP
TTTPVVT
TTXSVVX
XVTVTXTT
PPSV
VTTPV
(TSSSXxVv)
(P TTTTTV V )
(TS X X V V )
(rssxs)
(TX X TTvV )
(Prvv)
(TX X TTV P S )
(TSSXXVPS)
(P V P X TV V )
(P TTTTV V )
(TS X X TV V )
(TX X TTTV V )
(PvPS)
(P TTV V )
PVPXTTVV, PTTVPXVV, and
Note:"Theseitemshavethreecorrectsolutions:
PTVPXTVV.
bAnother
is: PVPXTVPS.
correctsolution
solutionsarein parentheses.
Grammatical
In this experimentyou haveto learn (stringsfrom 3 to 8 lettersin length / pairs
of letters).The composingletters are P, S, T, V, X. These(strings/ pairs) are
displayedon the next page.After the signal,you will have l0 minutesto study
them.
For both groups, the front page began with the sentence in upper case:
Do not turn the pagebefore the signal.
203
LEARNING
ANDABSTRACTIVENESS
IMPLICIT
When the signal was given, subjects studied the items (the grammatical
stringsof lettersor the bigrams)displayedon the secondpageof the booklet
for 10 minutes.The first booklets rverethen collectedby the experimenter.
Subjects\ryereorally requestedto take the secondbooklet out of the folder,
and to read the instructions on the front pâBe,which again beganwith the
sentence:
Do not turn the pagebeforethe signal.
"right" order.
The subjectswere asked to arrange the letter strings in the
Subjects in the string group were informed that the study items were
generatedby a set of rules, and correctnesswas definedas conformity with
theserules.The bigram group wastold that a correct letter string wasdefined
by the fact that eachpair of contiguouslettershad to belongto the previously
studied list. When the signal was given, subjects began to write their
responses.They wereurged to work quickly, but no time limit was imposed.
Results
Probabitity of Correct Solutions. In order to obtain a basis for assessment of the observedperformance,the probability of correct respondingwas
computedfor eachletter string asthe ratio of the number of correct solutions
(usually one, in a few casestwo or three) over the number of possible
arrangements.The mean probability was 0.013.The valuesfor the letter
strings of length 4 to 8 are reported in Table 4, first column.
For the string group, the meanproportion of correctanagramsolutions
was 0.113(s:0.137).As shownin Table 4, secondcolumn,therewas a
marked decreasein correct anagramsolutionswith lengtheningof the letter
TABLE4
of CorrectAnagramSolutionby Chance,and
Proportion
in the StringGroupandthe BigramGroup
Observed
Bigram GrouP
I*ngth
ChanceValue
StringGroup
4 (2)
6 (4)
7 (7)
8 (12)
0.083
0.039
0.006
0.006
0.001
0.230
0.307
0.117
0.100
0.052
0.095
0.063
0.010
0.010
0.007
Weightedmean
0.013
0.113
0'020
s (3)
Note: Proportions were calculatedfor each length of anagrams.Mean values
werecomputedafter weightingby the number of stringsrepresentativeof eachlength,
as shown in parentheses.
PACTEAU
GALLEGO,
204 PERRUCHET,
letter strings provide more
strings. This effect is not surprising, as long
opportunities for error.
lesswell than the
As expected,the bigram group performedconsiderably
of correct
proportiol
mean
The
string group (Table I, tniiA column).
on
dependent
closely
again
\ryas
solutions was 0.02 1s:b.043). Performance
chance
than
better
was
performance
the length of the tetàr strings.However,
probability of 0.031(1/25)of
for all lengrhs,a result which has a binàmial
occurring.r
of each of the 25
CorrelationalPattern. The proportion of occurrence
correctedfor the
then
subject,
each
possiblepairs of letterswas assesr.dfot
and Lewis'
Rgber
the
using
chance
likelihood of assemblingletters by
the
presents
5
Table
5)'
Table
to
note
correction f"t ri" tt.ptËduced in tlhe
frequencies
with
along
subjects,
over
raw and adjustedproportions averaged
of study and virtual bigrams'
were computedusing
The product-momeit correlationsfor thesevalues
for the 16 valid
computed
first
ffio different methods. Correlations were
computed their
Lewis
and
(Reberbigrams after averaging over subjects
We
information'
rank-order
used
correlationsin thi.-*Ëy,-althoughthey oniy
significance
statisticar
for
test
to
usedproduct_momenicorrelationsin order
in Table 6' left panel' An
of differences).The coefficient values appear
higher fot l!: bigram group
unpredicted result was that correlations were
1969,p' 158) show that the
than for the string group. T-tests(McNemat,
bigrams' t(13):6'15'
observed
the
difference,*.r.-r"i;rrin unt, for bàth
:3'!'
However'within
p:0'004'
(13)
p<0.001, and the iirtuat bigrams,
than for
bigrams
virtual
for
each group, the correlations *rr.'higher
to reach
failed
differences
the
observed bigraÀs- Although fairly lut-g.,
l..481'
(13):
:t'67;
group:
bigram
significance
lstringgroup: <f lt
basis,then averaged
Correlationswerealso computedon a within-subject
were higher for the
Correlations
over subjects(after r to z traniformation)'
bigrams only'
observed
for
group
bigram group than for the string
anagram solutions
I An anonymous reviewer of an earlier version of this paper noted that
maybeaffectedbylinguisticpreferences,',"ti,t'^"ycoincidewiththegrammaticalordering'In
ofaS ttod"nts in psychologywere askedto
order to test this typltn rir, an addition"igtoup
prior exposure to the study material' The
wiitrout
fit,
arrange the test strings as they saw
This result
was near chance(0.010vs. 0.013,respectively)'
proportion of grammùcal responses
bigram groups is
or
string
the
from
"ittUi*s
ensures that above-chanceperformance
study phase'
Àput"Ut. to their exposureto m3tgrial il the
on
traueno effect.Positive and negativeeffects
This doesnot entail that linguisti. pr.f.r.i""s
consider
unaffected'
perfonnance
leaving mean
specific strings may compensateone another,
45 subjectsgeneratedthe
thit-;;;.-Seien."*-:lthe
illustrate
to
strings
4-letter
two
the
suggestsa marked
which
whereas nJn. g"o"tated PVPS'
grammatical string itVV,
for the &lctter
responses
of
propoÀo,,
era.mlatical
preference effect. However, the mean
4)'
Table
in
level (0.0E3,as indicated
strings(llgl,i.e. o.oial is iose to chance
205
AND ABSTRACTIVENESS
IMPLICITLEARNING
TABLE5
of EachBigramin the StudyStringsand in the
Numbersof Occurrences
and Rawand
by the Grammar.
WholeSetof StringsGenerated
of EachBigram
of Occurrences
Proportion
Corrected
Occurrences of each
Bigram
Material
SS
xx
TT
vv
XS
sx
PS
xv
PX
XT
VP
TX
PV
TV
TS
PT
Study
Virtual
7
3
4
6
4
5
5
6
5
2
l0
2
4
6
5
4
l5
T7
23
23
6
t4
l4
l4
l0
l3
24
9
7
26
l4
13
Scores
StringGroup
Bigram Group
P
P'
0.056
0.052
0.123
0.086
0.031
0.031
0.028
0.049
0.014
0.030
0.037
0.057
0.044
0.072
0.027
0.043
0.022
0.027
0.049
0.050
-0.016
0.008
0.007
0.006
0.002
-0.020
0.004
- 0.002
-0.002
0.001
-0.029
0.007
0.045
0.041
0.099
0.071
0.033
0.037
0.026
0.048
0.014
0.046
0.056
0.044
0.046
0.084
0.030
0.044
0.011
0.016
0.023
0.035
- 0.014
0.015
0.005
0.00s
0.003
-0.003
0.024
-0.016
0.000
0.014
-0.027
0.008
P= Raw proportion of occurrences.
P : Correctedproportion of occurences.
Note: Correction for guessingwas computed using the Reber and Lewis formula (L977,
p. 342):p : (p - pù | (l - Pg), whereP is the proportion of actual occurrencesin subjects'
solutions,and Pg the proportion of occurrencesresulting from guessing.
TABLE6
Product-Moment Correlationsof the Frequencyof BigramsComposing
Subjects'Solutionswith the Frequencyof study Bigramsof one Hand,
and the Frequencyof the BigramsComposingthe Whole Set of Strings
the GrammarMay Generateon the Other Hand
r on Mean Scores
Study bigrams
Virtual bigrams
Means of Within-Subiectr
StringGroup
Bigram GrouP
StringGroup
Bigram GrouP
0.175
0.s47
0.458
0.730
0.14
0.43
0.26
0.43
group
Note: Coefficients\ilerecomputedusing two different methods(seetext) for string
group.
for
bigram
and
PAcTEAU
GALLEGo,
206 PERRUCHET,
(51) :2.52, p:0.014. Otherwise,the results(Table6, right panel)displaythe
samegrtt.t"l pattern as above.Strikingly, the correlationswere significantly
higheiwith virtual bigramsthan with observedbigrams,for both the string
gtorrp,t(25):10.61,p<0.001,and the bigramgroup,t(26):3.40,p:0.00'2.
The fact that correlationsof interestwere generallyhigher for the bigram
group than for the string group, especiallywhen the data involving observed
stringswere examined,\ryasnot anticipated.This differenceclearly indicates
that the performance of the string group was less dependenton bigram
frequencythan wasbigram group performance.The abstractionistinterpretation would be that perforrnancein the string group was also affectedby the
internal representationof higher-level rules. However, this result is also
congruentwith a generalframework positing that implicit learningonly gives
u.rÀr to specificfeaturesof the stimuli. In the string group, frequencyof
occurrencemay competewith other factors in determining the strength of
memory tracesof the bigrams,in particular with the position of the bigrams
analysisof resultsdisplayedin Table 5
within the strings.A case-by-case
supportsthis interpretation.For instance,bigram VV had the best scorein
thè string group, although it was initially displayedlessoften than SS and
VP. This may be accountedfor by positional factors: all VV bigramsare
terminal, and thus highly salient, whereasall SS and VP bigrams are
internal.
Discussion
As in Reber and Lewis (1977), the frequencyof bigrams generatedby
subjects trained on grammatical letter strings correlated better with the
frequencyof the bigramsmaking up the whole setof stringsthe grammarcan
genèratethan with the frequencyof the bigramsinitially displayed.However,
àn identical pattern was observed when subjects were first exposed to
individual pairs of letters constitutiveof grammaticalstrings,a condition
precludingthe abstractionof high-levelrules. This latter finding rules out
Reber's (1989) contention that obtaining this pattern of correlationsin
subjectstrained in standardconditions providesevidencefor the abstraction
of the deepstructure of the grammar usedto generatethe strings.
One poiential problemwith our resultsis that the pattern of correlations,
although reflectingthe samegeneraltrend as the one reportedby Reber and
Lewis 1tVll1, exhibits much lesscontrast.Correlationsof performanceof
subjectsstudyinggrammaticalletter stringswith observedand virtual bigram
in our experiment,whereasthe
frequencier*.tè 0.17and 0.55,respectively,
,ori..ponding valuesin the Reberand Lewisstudywere0.04and 0.72.There
is emfirical èvidencefor attributing these differencesto the differential
urno.rntof subjecttraining. In the Reber and Lewis study, in which subjects
had four learning sessions,correlationsof performancewith virtual bigram
ANDABSTRACTIVENESS
LEARNING
207
IMPLICIT
frequencieswere0.44,0.58,0.68, and0.77, respectively,for SessionsI to 4
(Reber & Lewis, 1977,Table 7. Correspondingdata for observedbigrams
were not reported). Our subjectshad only one learning session,and their
mean performancelevel (0.1I correct solutions)was notably lower than the
one reportedby Reberand Lewis (overall mean:0.46correct solutions),even
on their first session(0.31correctsolutions).On this basis,it could be argued
that our results only pertain to the early phase of learning, and that
abstraction of deepstructure occursat a later stage.
This objection calls for several comments. (l) Differencesin level of
learning could be less marked than the proportion of correct solutions
suggests.In the Reber and Lewis procedure,subjectshad to reorder a set of
shuffied cards, each containing one letter; thus they could not make any
mistakesregarding the number or nature of the letters in their solution. In
addition, a letter was placed in its proper location on three-quartersof the
trials. Thus part of the differencein performancemay be due to the fact that
the Reber and Lewis subjectshad their responsesconstrainedto a greater
extent than ours. (2) Our learningconditions in fact closelyparalleledthose
usedin earlier studies,which claim to provide evidencefor implicit abstraction. Subjectsstudied the letter strings for ten minutes; as a casein point,
Reber,Kassin,Lewis,and Cantor (1980,experimentl) displayedtheir study
list, which was generatedby the samegrammar as the presentone, for only
sevenminutes.(3) As arguedin the introduction, extendingthe duration of
Thus the data collected after
learning has severalnegativeconsequences.
intensivetraining in the Reber and Lewis experimentmay not be relevantto
implicit learning. Further and more damaging to their arguments, it is
doubtful that the correlational pattern obtained after intensivetraining has
underlyingperformance,ascorrelations
any significanceas regardsprocesses
approach fixed values as subjectssolve more and more anagrams.Thus,
overall, differencesin the amount of training between Reber and Lewis'
studiesand ours do not seriouslyundermineour conclusions.
Another possiblereservationconcernsthe type of evidenceprovidedin the
presentstudy. Showingthat the Reber and Lewis correlationalresultsare
flawed doesnot entail that strongercorrelationswith bigrams composing
studythan virtual stringswould be obtainedin more valid conditions;hence,
our study doesnot indicate that subjectsare more sensitiveto study than to
the virtual strings.rWefully agreethat we provide only negative,and hence
limited, evidence.However,as statedin the introduction,most of the biases
to which we draw attention are intrinsic componentsof the Reberand Lewis
procedure.Thus carrying out new experimentspatternedafter the Reberand
Lewis procedureafter eliminating its biasesconstitutesa dead end. We are
not challengrngthe value of anagramtechniqueas a meansof investigating
knowledgeacquired in the artificial grammar setting. Rather, our point is
208 PERRUcHET,
GALLEGo,
PAcTEAU
that the techniqueis ill suited for the specificusethat was the main focus of
the presentarticle.
SUMMARY
A N DC O N C L U S I O N
Earlier work on artificial grammar learning has shown that specific and
consciousknowledgeregarding,for instance,valid bigramsor trigiams could
be sufficientto accountfor grammaticalityjudgements(Druhan & Muthr*r,
1989;Dulany et al., 1984;Perruchet& pacteau,1990).Thesefindingsdo not
exclude the fact, however, that people can acquire some unconscious
representationof the underlying structure of the grammar. In his recent
theoreticaloverviewon implicit learning, Reber (1989)puts forward two
argumentssupportingthis assumption.The presentarticle showsthat both
argumentsare groundedon experimentaldata that can be accountedfor in
other ways.
what is challengedin this and earlier works (perruchet et al., 1990;
Perruchet& Pacteau,1990,l99l) may be definedasimplicit abstraction,that
is, the unconsciousformation of new abstractstructures.Our position is that
in current experimental settings, implicit learning conditiôns may only
generatespecificknowledge.We take it for grantedthat peoplemay abstract
rules from factual information. However, abstraction may be fin[ed to the
use of logical reasoningand analytic modesof processing.
This contentiondoesnot minimizethe importanceof implicit learningin
adaptivebehaviour.Restrictingimplicit learningto the acquisitionof speàfic
knowledge at the expenseof abstract structures or rules should not be
construedas a major limitation on its explanatorypower. Thereis growing
evidencein the fields of categorizationand problem-solvingfor models
positing the primacy of the specificknowledgein the formation of abstract
structures.Most of the data that seemedat one time to be straightforward
support for rule, schema,or prototype abstraction have been reinterpreted
without assumingthat abstractiveabilitiesare calledinto play. For instance,
it has been repeatedlyshown that subjectscan classify previously unseen
prototypesmore accuratelythan exemplarson which they were trained (e.g.
Posner& Keele,1968;parenthetically,this bearsstriking resemblance
to the
correlationaldata in this article,as Reber,19g9,noted).Medin and shaffer
(1978) show that a mathematicalmodel based on their exemplar-based
theory of categorizationcan predict the Posnerand Keeleresultsas well as a
prototype theory. Other related works aimed at illustrating that a large
amount of adaptive behaviour derives from the acquisition and use ôf
specificknowledgeare describedin the Medin and RoJs (1990)review.
IMPLICIT
LEARNING
ANDABSTRACTIVENESS
209
RE F E R
E NC E S
-
"
.
.
'
Berry, D.C. & Broadbent,D.E. (1988).Interactivetasksand the implicit-explicitdistinction.
British Journal of Psychology,79,251172.
Broadbent,D.E. (in press). Recall,recognition,and implicit knowledge.In W. Kessen,A.
Ortony, & F.I.M. Craik (Eds.), Festschriftfor George Mandler (provisional title)
(pp. 125-134).Hillsdale,NJ: LawrenceErlbaum Associates,Inc.
Broadbent,D.E., Fitzgerald,P., & Broadbent,M.H.P. (1986).Implicit and explicitknowledge
in the control of complex systems.BritishJournal of Psychology,77,33-50.
Brooks, D.N. (1978). Nonanalytic conceptformation and memory for instances.In E. Rosch
& B.B. Lloyd (Eds.), Cognition and categorization.Hillsdale, NJ: Lawrence Erlbaum
Associates.
Inc.
Druhan, B.B. & Mathews, R.C. (1989).THIYOS: A classifiersystemmodel of implicit
knowledgeof artificial grammars.Paper presentedat the annual meetingof the Cognitive
ScienceSociety,Ann Arbor, Michigan.
Dulany, D.8., Carlson,A., & Dewey,G.I. (1984). A caseof syntacticallearningandjudgment :
How consciousand how abstract?Journal of ExperimentalPsychology:Gmeral, I13, 541555.
Estes,W.K. (1986). Array modelsfor categorylearning. CognitivePsychology,/8,50O-549.
Hayes,N.A. & Broadbent,D.E. (1988).Two modesof learningfor interactivetasks.Cogaition,
28,249-276.
Holland, J.H., Holyoak, K.J., Nisbett, R.E., & Thagard,P.R. (1986).Induction:Processes
of
inference,learning,and discovery.Cambridge,MA: The M.I.T. Press.
Jacoby, L.L. & Brooks, L.R. (1984). Non-analytic cognition: Memory, perception, and
concept learning. In G. Bower (Ed.), The psychologyof learningand motivation, Vol. 18
(pp. l-af. New York, AcademicPress.
Lewicki, P., Czyznwska,M., & Hoffman, H. (1987). Unconscious acquisition of complex
procedural knowledge. Journal of Experimental Psychology:Lcarning, Memory and Cognition,I3,523-530.
Lewicki, P., Hill, T., & Bizot, E. (1988). Acquisition of proceduralknowledgeabout a pattern
of stimuli that cannot be articulated,.CognitivePsychology,20,24-37.
Mathews, R.C. (1990).Abstractnessof implicit grammar knowledge:Commentson Perruchet
and Pacteau'sanalysisof syntheticgrammar learning.Journal of ExperimentalPsychology:
General,I19, 412416.
Mathews, R.C., Buss, R.R., Stanley,W.8., Blanchard-Fields,F., Cho, J.R., & Druhan,
B. (1989). Role of implicit and explicit processesin learning from examples:A synergistic
effæt. Journal of ExperimentalPsychology:Learning, Memory, otd Cognition, /5, 1083I100.
McAndrews, M.P. & Moscovitch, M. (1985). Rule-basedand exemplar-based
classificationin
artificial grammar learning. Memory and Cognition, 13,469475.
McNemar, Q. (1969). Psychologicalstatistics.New York: J. Wiley and Sons(4th edition).
McKelvie, S.J. (1987). Learningand awarenessin the Hebb digits task. TheJournalof General
Psychology,I I 4, 75-88.
Medin, D.L. & Shaffer,M.M. (1990).A contexttheoryof classification
learning.Psychological
Review,85, 207-238.
Medin, D.L. & Ross,B.H. (1990).The specificcharacterof abstractthought:Categorization,
problem solving, and induction. In R.J. Sternberg(Ed.), Thepsychologyof hwnan intelligence,Vol. 5 (pp. 189-223).Hillsdale, NJ: LawrenceErlbaum Associates,Inc.
Millward, R.B. & Reber,A.S. (1972).Probabilitylearning:Contingent--event
sequences
with
lags.AmericanJournal of Psychology,8J, 8l-98.
Nissen,M.J., & Bullemer,P. (1987).Attentional requirementsof learning: Evidencefrom
performancemeasures.CognitivePsychology,J,9,l-32.
210
PERRUcHET,GALLEGo, PAcTEAU
Perruchet, P., Gallego, J., & Savy, I. (1990). A critical reappraisal of the evidence for
unconsciousabstractionof deterministicrulesin complexexperimentalsituations.Cognitive
Psychology,22, 493-516.
Perruchet,P. & Pacteau,C. (1990). Syntheticgrarnmar learning: Implicit rule abstraction or
explicit fragmentaryknowledge?fournal of ExperimentalPsychology:General,119,2&-275.
Perruchet, P. & Pacteau,C. (1991). The implicit acquisition of abstract knowledge about
artificial grammar: Some methodological and conceptual issues.Journal of Experimental
Psychology: General,I 20, 112-116.
Posner,M.I. & Keele, S.W. (1968). On the genesisof abstract ideas.Journal of Experimental
Psychology,77, 353-363.
Reber, A.S. (1967). Implicit learning of artificial grarnmars.Journol of Verbal kaming and
YerbalBehavior,6, 855-863.
Reber,A.S. (1989). Implicit learningand tacit knowledge.Journalof ExperimentalPsychology:
General,I18,219-235.
Reber,A.S., Allen, R., & Regan,S. (1985) Syntacticallearningandjudgment, still unconscious
and still abstract: Comment on Dulany, Carlson, and Dewey. Journal of Experimental
Psychology
: General,I 14, 17-24.
Reber,4.S., Kassin,S.M., Lewis,S.,& Cantor, G. (1980). On the relationshipbetweenimplicit
and explicit modes in the learning of a complex rule structure. Journal of Experimental
Psychology: Human l*arning and Memory, 6, 492-502.
Reber,4.S., & Lewis, S. (1977). Implicit learning:An analysisof the form and structure of a
body of tacit knowledge.Cognition,5, 33l-36 I .
Sanderson,P.M. (1989). Verbalizableknowledgeand skilled task performance:Association,
dissociation,and mental models.Journal of ExperimentalPsychology:Learning, Memory,
and Cognition,I 5, 729-747.
Stanley,W.8., Mathews,R.C., Buss,R.R., & Kotler-Cope,S. (1989).Insight without a$'areness:on the interaction of verbalization, instruction and practice in a simulated process
control task. The QuarterlyJournal of ExperimentalPsychology,4lA, 553-577.
Revisedmanuscriptreceived7 January 1991